Biodius / Research

Negative evidence belongs in the record.

Technical Report / Preprint · Not peer reviewed

Falsification-First AI for Biological Target Assessment: Lessons from Four Prospectively Frozen Experiments

Dr. Hernani Costa · ORCID 0000-0002-6813-4641
Hernani Holding B.V.

Version 1.1 · 1 October 2026 · CC BY 4.0

Abstract

AI systems can produce plausible biological stories faster than people can test them. This report asks whether a research workflow can make unsupported conclusions harder to keep. It reports four prospectively bounded experiments and retains the negative, discordant and not-evaluable outcomes they produced. It does not validate a therapeutic target, clinical intervention, diagnostic, safety claim or general-purpose drug-discovery system.

Four prospectively bounded experiments

Experiment 01

Heavier architecture

A prospective pulmonary-fibrosis benchmark found that a heavier scientific-agent architecture did not demonstrate superior useful biological coverage over a strong agentic research workflow. It provided safeguards around leakage, provenance, lineage and reproducibility; complexity still had to earn its place.

Experiment 02

Lean Scientific Core

A smaller scientific trust layer transferred once to inflammatory bowel disease, preserving useful coverage and adding decision-relevant annotations around provenance, independence, boundaries and contradictions. This was software safeguard evidence, not biological evidence.

Experiment 03

Causal-chain challenge

A plausible ulcerative-colitis hypothesis survived initial adversarial review. A prospectively frozen public-human-genetics adjudication could not establish the required HNF4A → SLC26A3 mediation bridge from lawful public aggregate data. The hypothesis was not promoted: NOT_EVALUABLE does not mean false.

Experiment 04

TTPAL independent challenge

A gut-resident T-cell discovery signal reached PP4 up to about 0.992. The independent peripheral-blood challenge did not support it. The signal remained single-dataset-only; the difference in tissue context means this does not rule out a gut-resident effect.

What the evidence supports

Frozen questions, evidence boundaries and independent challenges can retain useful negative, discordant and not-evaluable outcomes.
In the recorded architecture comparison, heavier architecture had no measured discipline advantage over the strong baseline.
The TTPAL discovery result was a reason for challenge, not target promotion.

What it does not support

The evidence does not establish a therapeutic, clinical, diagnostic or safety conclusion. It does not establish a general ranking of research architectures or universal AI performance. It does not show that the causal chain is false. It does not call TTPAL irrelevant. It does not use a blood null to rule out a gut-resident effect.

Limitations

The report is limited to four recorded experiments. The architecture comparison had one evaluator; Lean Core reports one transfer; the causal-chain challenge remains data-limited; and the TTPAL discovery and independent challenge differ in tissue context and lineage.

Cite this work

Costa, Hernani. (2026). Falsification-First AI for Biological Target Assessment: Lessons from Four Prospectively Frozen Experiments (Version 1.1, Technical Report / Preprint). Zenodo. https://doi.org/10.5281/zenodo.23100159

Version 1.1 · DOI 10.5281/zenodo.23100159
All versions · DOI 10.5281/zenodo.23057911